Toward Efficient and Robust Large-Scale Structure-from-Motion Systems Public Deposited

Downloadable Content

Download PDF
Last Modified
  • March 19, 2019
  • Heinly, Jared
    • Affiliation: College of Arts and Sciences, Department of Computer Science
  • The ever-increasing number of images that are uploaded and shared on the Internet has recently been leveraged by computer vision researchers to extract 3D information about the content seen in these images. One key mechanism to extract this information is structure-from-motion, which is the process of recovering the 3D geometry (structure) of a scene via a set of images from different viewpoints (camera motion). However, when dealing with crowdsourced datasets comprised of tens or hundreds of millions of images, the magnitude and diversity of the imagery poses challenges such as robustness, scalability, completeness, and correctness for existing structure-from-motion systems. This dissertation focuses on these challenges and demonstrates practical methods to address the problems of data association and verification within structure-from-motion systems. Data association within structure-from-motion systems consists of the discovery of pairwise image overlap within the input dataset. In order to perform this discovery, previous systems assumed that information about every image in the input dataset could be stored in memory, which is prohibitive for large-scale photo collections. To address this issue, we propose a novel streaming-based framework for the discovery of related sets of images, and demonstrate our approach on a crowdsourced dataset containing 100 million images from all around the world. Results illustrate that our streaming-based approach does not compromise model completeness, but achieves unprecedented levels of efficiency and scalability. The verification of individual data associations is difficult to perform during the process of structure-from-motion, as standard methods have limited scope when determining image overlap. Therefore, it is possible for erroneous associations to form, especially when there are symmetric, repetitive, or duplicate structures which can be incorrectly associated with each other. The consequences of these errors are incorrectly placed cameras and scene geometry within the 3D reconstruction. We present two methods that can detect these local inconsistencies and successfully resolve them into a globally consistent 3D model. In our evaluation, we show that our techniques are efficient, are robust to a variety of scenes, and outperform existing approaches.
Date of publication
Resource type
Rights statement
  • In Copyright
  • Berg, Alexander
  • Dunn, Enrique
  • Niethammer, Marc
  • Frahm, Jan-Michael
  • Agarwal, Sameer
  • Doctor of Philosophy
Degree granting institution
  • University of North Carolina at Chapel Hill Graduate School
Graduation year
  • 2015
Place of publication
  • Chapel Hill, NC
  • There are no restrictions to this item.

This work has no parents.